ABSTRACT
Robustness is a characteristic of regulatory pathways to ensure signal consistency in light of environmental changes or genetic polymorphisms. The retinoic acid (RA) pathway is a central developmental and tissue homeostasis regulatory signal, strongly dependent on nutritional sources of retinoids and affected by exogenous chemicals. We performed transient physiological RA signaling disturbances during embryogenesis followed by kinetic transcriptomic and high-throughput qPCR analysis of the recovery. Unbiased pattern analysis identified the RA metabolic network as the main regulated module aimed at achieving signaling robustness. We used a principal trajectory-based analysis of the clutch-dependent variability and organized the results into a robustness efficiency matrix comparing the RA feedback regulation and hox gene expression (RA targets). We found the feedback autoregulation to be sensitive to the direction of the RA perturbation: RA knockdown exhibited an upper response threshold, whereas RA addition did not activate a feedback response below a minimum threshold. These results demonstrate an asymmetric capacity for robust feedback control of the RA signal during early embryogenesis, probably based on genetic polymorphisms, likely a significant contributor to the manifestation of developmental defects.
Introduction
Retinoic acid (RA) is a central regulatory signaling pathway active during embryogenesis and adult tissue homeostasis (le Maire and Bourguet, 2014; Metzler and Sandell, 2016). RA is synthesized from vitamin A (retinol) or other retinoids or carotenoids obtained from the diet (Blaner et al., 2016; Fainsod and Kot-Leibovich, 2018; Ghyselinck and Duester, 2019; Kedishvili, 2016). During embryogenesis, excessive or reduced RA signaling induces a wide array of embryonic malformations (Corcoran et al., 2002; Durston et al., 1989; Hollemann et al., 1998; Kessel and Gruss, 1991; Kot-Leibovich and Fainsod, 2009; Marshall et al., 1992; Papalopulu et al., 1991). Disorders linked to reduced RA signaling include vitamin A deficiency syndrome, DiGeorge/VeloCardioFacial syndrome, Fetal Alcohol Spectrum Disorder, Congenital Heart Disease, neural tube defects, and multiple types of cancer (Hartomo et al., 2015; Kim et al., 2005; Kot-Leibovich and Fainsod, 2009; Pangilinan et al., 2014; See et al., 2008; Urbizu et al., 2013). Therefore, RA levels are tightly regulated throughout life, controlling quantitative, spatial, and temporal expression and activity of RA biosynthetic and metabolizing enzymes (Blaner, 2019; Ghyselinck and Duester, 2019; Kedishvili, 2016).
RA biosynthesis involves two sequential oxidation steps: first, mainly alcohol dehydrogenases (ADH) or short-chain dehydrogenase/reductases (SDR) oxidize vitamin A (retinol, ROL) to retinaldehyde (RAL), followed by the retinaldehyde dehydrogenase (RALDH) catalyzed oxidation of RAL to RA (Ghyselinck and Duester, 2019; Kedishvili, 2016; Metzler and Sandell, 2016). ROL and RAL can also be produced from retinyl esters or ß-carotene from food sources (Blaner, 2019). RA availability is further affected by ROL, RAL, and RA binding proteins (Napoli, 2017). In gastrula vertebrate embryos, RA signaling is triggered by the expression of RALDH2 (ALDH1A2) completing the biosynthesis of RA (Begemann et al., 2001; Chen et al., 2001; Grandel et al., 2002; Niederreither et al., 1999). Substrate (RAL) availability for the RALDH enzymes is controlled by RAL reductases (Adams et al., 2014; Billings et al., 2013; Feng et al., 2010; Porté et al., 2013; Shabtai and Fainsod, 2018). Besides the temporal and spatial regulation of RA metabolic network components, other RA network components, including the RA receptors (RAR and RXR) and retinoid-binding proteins (Cui et al., 2003; Janesick et al., 2015; Lohnes et al., 1995) are under tight regulation. Then, RA signaling is controlled through its regulated biosynthesis and degradation, as well as restricted spatiotemporal expression of the metabolic and gene-regulatory components, which together are expected to provide robustness to RA signaling. The maternal nutritional status and environmental exposure to chemicals such as alcoholic beverages (ethanol) and others can also affect the RA signal (Paganelli et al., 2010; Shabtai et al., 2018). These observations point to the close interaction of RA signaling and the environment and the necessity to adapt the RA signaling network to nutritional changes and insults. This adaptation and maintenance of normal signaling levels under changing conditions is termed robustness (Eldar et al., 2004; Nijhout et al., 2019).
A deeper understanding of the RA signaling pathway during embryogenesis is required to elucidate its multiple regulatory roles and the regulation of the signaling robustness in the presence of disturbances. Commonly, the RA pathway is studied by increasing this signal (Durston et al., 1989; Kessel, 1992; Sive et al., 1990), or knockdown studies take advantage of inhibitors, inverse agonists, or enzymatic degradation (Hollemann et al., 1998; Janesick et al., 2013; Janesick et al., 2014; Kot-Leibovich and Fainsod, 2009). In multiple RA knockdown studies, the developmental malformations observed were milder than expected (Blumberg et al., 1997; Hollemann et al., 1998; Janesick et al., 2014; Koide et al., 2001; Shabtai et al., 2018; Sharpe and Goldstone, 1997) suggesting a compensatory mechanism. To characterize the robustness of RA signaling during early embryogenesis, we devised clearly defined and transient RA manipulations in the physiological range that are terminated during early gastrula, and monitored by RNAseq, qPCR, and phenotypic analysis for several hours during recovery. These results demonstrated high RA signaling robustness resulting in relatively small transcriptome-wide changes. Unbiased detailed transcriptomic analysis revealed that RA metabolic and signaling network components exhibited expression changes dependent on the direction of RA manipulation, suggesting a mechanistic explanation for the robustness observed. Analysis of multiple embryo clutches (biological replicates) revealed differences in their individual robustness response to enhanced or reduced RA signaling. These results exposed potential consequences of the underlying genetic differences between clutches. Our results suggest an asymmetric capacity for RA signaling robustness in the early embryo, likely contributing to the human developmental defects that arise due to RA imbalance, most often, reduced Vitamin A levels during early pregnancy.
Results
Physiological RA manipulation uncovers signaling robustness
To manipulate RA signaling in early embryos, we routinely employ two approaches to knockdown this ligand (Fig. 1A). We either partially inhibit the oxidation of RAL to RA with RALDH inhibitors like 4-diethylaminobenzaldehyde (DEAB) (Shabtai et al., 2016), or we overexpress the RA hydroxylase, CYP26A1 (Hollemann et al., 1998; Yelin et al., 2005), to render RA inactive (Topletz et al., 2015). These treatments result in unexpectedly mild developmental defects (Figs. 1B-D). To induce severe phenotypes by RA loss-of-function we combine knockdown treatments targeting the metabolism at different steps (Fig. 1A, E) supporting the robustness of RA signaling.
To study the RA network robustness in the physiological range we empirically determined the RA concentrations to use by testing concentrations (1 nM - 1 μM) spanning the reported physiological content of X. laevis early embryos (100-150 nM all-trans RA) (Durston et al., 1989; Kraft and Juchau, 1992; Kraft et al., 1995; Sive et al., 1990). We analyzed gene expression of RA metabolic enzymes and RA targets during early (st. 10.25) and late gastrula (st.12) (Nieuwkoop and Faber, 1967). Genes positively regulated by RA at both stages, hoxb1, cyp26a1, and dhrs3 exhibited dose-dependent responses (Fig. 2A-C). The genes raldh2 and rdh10, are downregulated by most RA doses (Fig. 2D, F), while raldh3 (aldh1a3) is upregulated earlier on and is repressed later on (Fig. 2E). Importantly, RA concentrations as low as 10 nM had a significant effect on gene expression.
We performed transient pharmacological RA manipulations within the physiological range in early Xenopus embryos. The perturbations were terminated by washing the treatments and then the embryos were monitored during the recovery period. Embryos were treated with RA (10 nM) for 2 hours starting from late blastula (st. 9.5) and washed during early gastrula (st. 10.25). At different time points during the post-wash recovery period, samples were collected to perform kinetic analysis of the changes in gene expression by qPCR (Fig. 3). To monitor the changes in RA signaling levels by gene expression, two well-characterized RA-regulated genes, hoxb1 and hoxb4, were analyzed. Both genes showed that at the time of treatment washing (t0), they were upregulated when compared to control sibling embryos (Fig. 3A). Two hours into the recovery, both RA target genes were back to control expression levels. To begin understanding the robustness of RA signaling and the dynamic regulation of RA metabolism we studied the expression of dhrs3, which preferentially reduces RAL to ROL to attenuate RA biosynthesis (Feng et al., 2010), cyp26a1, which targets RA for degradation, and raldh2 that produces RA from retinaldehyde (Shabtai et al., 2016). Both negative RA regulators, dhrs3 and cyp26a1 are upregulated at t0 by increased RA (Fig. 3B) and return to normal expression levels after 4-hour recovery. As expected, raldh2 exhibited an RA-promoted downregulation at t0 (Fig. 3B), returning to normal levels after 4 hours. These results indicate that the gene expression of RA metabolic enzymes, auto-regulated by RA (Fig. 2), is shifted to achieve normal RA levels in the face of external RA perturbation.
A complementary study was performed by inhibiting RA biosynthesis with DEAB (Fig. 3C,D). All genes studied exhibited fluctuations during the recovery period. Also, in this case, hoxd1 and hoxb1 reached almost normal levels at an earlier stage than the genes encoding RA metabolic components (Fig. 3C). Genes encoding anabolic enzymes, e.g., raldh2, were upregulated, and catabolic components, e.g., cyp26a1, were downregulated (Fig. 3D).
Physiological RA perturbations uncover signaling robustness at the transcriptomic scale
To gain a better understanding of the RA signaling robustness, we performed a kinetic transcriptomic analysis. To optimize the kinetic study, parameters such as developmental stages and timing between samples were empirically tested. We treated the embryos with RA (10 nM) or DEAB (50 μM) for 2 hours from late blastula (st. 9.5) to early gastrula (st. 10.25) and collected samples every 1.5 hours after terminating the treatment (Fig. 4A). For each biological repeat, all treatments, controls, and time points were collected from a single clutch (eggs) of a single female. The efficiency of the treatments and quality of the RNA samples was first monitored by qPCR for changes in hox gene expression as a readout of the RA signal levels. Only those biological repeats where both treatments exhibited the expected upregulation (RA) or downregulation (DEAB) in hox gene expression were selected for RNA-seq.
We analyzed the time series transcriptomic data set for differentially expressed genes using a two-way ANOVA (t=0, 1.5, 3, 4.5 hours; treatments: RA, DEAB, control; n=6 biological replicates). Principal Component Analysis (PCA) of the gene expression variation showed separation along the first principal component (PC1) which corresponds to the developmental stage (Fig. 4B). Surprisingly, RA manipulated samples clustered with the control samples of the same developmental stage. Normal transcriptomic changes as a result of progression through embryogenesis appear to be the dominant variable distinguishing the samples irrespective of treatment (Fig. 4B). The second component (PC2) separated the sample groups at intermediate time points (1.5 and 3h) from those at 0 and 4.5h time points, indicating a transient differential expression shift as the next most dominant pattern in the data. The effects of RA and DEAB on the transcriptome was not readily apparent in the next seven principal components. The PC8 showed some separation of the DEAB group from the RA and Control groups (Fig. 4C), whereas PC10 showed some separation of the RA treatment group from Control and DEAB treatments (Fig. 4D). The top-ranked genes along PC8 and PC10 showed distinct dynamic patterns across the treatments, whereas top-ranked genes along PC1 and PC2 largely corresponded to in-common dynamic changes over time (Fig. 4D; Supplemental Fig. S1). The overall magnitude of induced changes in gene expression in response to RA or DEAB appears to be less than the normal transcriptome changes occurring during early developmental stages, showing that RA signaling robustness possibly dampens the gene expression changes otherwise induced by abnormal RA levels. These results are consistent with the response of a robust system that functions to limit the gene expression changes in the majority of the transcriptome.
Dynamic pattern analysis reveals the molecular genetic mechanisms underlying the RA robustness response
For characterization of the molecular transcriptomic mechanism of the RA robustness response, we analyzed the clutch-averaged data using an unbiased dynamic pattern analysis approach to categorize the gene expression profiles along all possible discretized patterns (Kuttippurathu et al., 2016). Our analysis revealed a total of 4693 significantly differentially expressed genes with greater than 2-fold changes over time (compared to t0 within each treatment group; multiple testing corrected q-value < 0.05). For each gene, a pattern based on the direction of expression change above a 2-fold threshold, or no change at each of the three recovery time points relative to t0 was determined. Such an exhaustive approach allows us to enumerate the dominant as well as subtle patterns in the data overcoming limitations of conventional cluster analysis that could miss or mask the smaller groups of genes with distinct expression profiles over time. Out of the 27 (3*3*3) possible dynamic patterns, only 10 patterns were exhibited by genes in the transcriptome in at least one of the three experimental groups (RA, DEAB, Control) (Fig. 5A). Of these, only 5 patterns were exhibited by a substantial number of genes (>100) in any experimental group, corresponding to up- or downregulation at later time points (3h and 4.5h) (Fig. 5A). At the 2-fold threshold, there were no genes that showed downregulation first and then upregulation and vice versa. There were no significant differences in the gene counts between RA or DEAB and control groups (two-tailed Z test, p > 0.05). Gene Ontology enrichment analysis of the control group highlighted several functional annotations including development, cell cycle, morphogenesis, RA biosynthesis, gastrulation, and others. (Fig. 5B; Supplemental Table S1). The patterns with relatively fewer genes (<100) correspond to transient up- or downregulation at 3h that is normalized by 4.5h, as well as progressive downregulation.
The pattern analysis (Fig. 5) shows that at the 2-fold threshold, there are no distinctive dynamic gene expression patterns that occur only in response to the RA or DEAB treatments. Also, the number of genes changing per pattern is similar between the two treatments and control suggesting that the physiologically relevant concentrations of RA and DEAB modulate the RA network components at a scale resembling the changes in controls. For instance, only a subset of genes that showed late downregulation (first row, Fig. 5A) was common across all the three groups (772 genes, Fig. 5C). Several genes showed similar late downregulation between the treatments or between one of the treatments and the Control group (Fig. 5C). In addition, many genes showed a late downregulation pattern only in one of the samples (Fig. 5C). While the Venn diagram analysis (Fig. 5C) revealed extensive overlap between both treatments and Control samples for each pattern, the overlap across patterns is not immediately clear from this analysis.
To exhaustively compare the RA and DEAB treatment groups for their effect on gene expression, we adapted our unbiased approach named COMPACT for analyzing time-series differential transcriptomic profiles across multiple experimental conditions (Kuttippurathu et al., 2016). For statistically significant changing genes within each treatment group (two-way ANOVA; q<0.05), we constructed discrete patterns based on differential gene expression between treatment (RA or DEAB) and controls (81 theoretically possible patterns, with up-, down- and no-regulation at t=0, 1.5, 3 and 4.5 h). Only the patterns with at least one gene in either comparison were included, yielding 32 distinct patterns (Fig. 6A). Finally, we intersected the two distributions to create a matrix of comparative patterns where each element corresponds to a distinct pair of patterns corresponding to RA vs. control and DEAB vs. control (Fig. 6B). The discrete patterns were based on a 1.3-fold average difference between RA (or DEAB) and control groups at each of the four time points. The effect size threshold was chosen at a lower level than 2-fold as the differential expression analysis revealed that the RA/DEAB perturbations were leading to a smaller magnitude of changes at each time point, as compared to the larger changes occurring normally over time, additional suggestive evidence of robustness.
The COMPACT matrix (Fig. 6B, Extended Data 1) shows that the majority of the genes sensitive to RA manipulation exhibited sensitivity to only one direction of the RA perturbation. A total of 193 genes only responded to the addition of RA (middle row), whereas 224 genes were only responsive to RA knockdown (DEAB; middle column). Of the genes responsive to both RA and DEAB perturbations, 88 genes showed up- or downregulation irrespective of RA or DEAB treatment (quadrants a and d; Fig. 6B). A set of 48 genes showed opposite transcriptional outcomes to RA increase versus RA decrease (quadrants b and c; Fig. 6B). Interestingly, the set of genes that showed upregulation by exogenous RA addition and downregulation by DEAB (quadrant b) contained multiple genes involved in RA metabolism (Fig. 6C).
We mapped the expression changes in Fig. 6B to the RA signaling and metabolic network (Fig. 6D,E). Notably, genes encoding proteins involved in suppressing RA levels (cyp26a1, dhrs3) were upregulated and the genes encoding enzymes for RA production (aldh1a2, aldh1a3, and rdh10) were downregulated in response to transient RA increase. These genes showed opposite regulation in response to DEAB treatment (Fig. 6E). As an independent evaluation, we analyzed the differential expression time series data using another unbiased approach, weighted gene correlation network analysis (WGCNA) (Langfelder and Horvath, 2008). A WGCNA gene expression module with the highest correlation with treatment contained a similar gene set as shown in Fig. 6C, supporting the COMPACT analysis (green module, Supplemental Table S2; Extended Data 2). Taken together, our results support a mechanism to maintain RA signaling homeostasis to counter the effects of exogenous perturbations and prevent teratogenic effects by a transcriptional feedback control system.
Clutch-wise heterogeneity uncovers alternative mechanisms for RA robustness
The analysis above shows changes in RA network components and identified the main RA network components responding during early gastrulation (Fig. 6). Interestingly, the PCA analysis suggested clutch-specific heterogeneity in gene expression changes where all samples from the same clutch tended to cluster together but the different clutches separated slightly from each other (Fig. 4B,C; Supplemental Fig. S2). For these reasons, we analyzed the transcriptional changes in the RA metabolic network within each clutch. We generated an additional set of samples to test the dynamic expression changes in RA metabolic network genes using high-throughput real-time PCR (HT-qPCR; Fluidigm Biomark; Supplemental Table S2) to validate RNAseq results. These six additional clutches (biological repeats labeled G-L) were treated with RA or DEAB following the same experimental design (Fig. 4A) used for the RNAseq study (clutches A-F). We observed a similar heterogeneity between clutches by HT-qPCR to the clutch-to-clutch variation in the RNAseq data (Supplemental Figs. S2 and S3).
We ordered the 12 biological repeats (clutches) according to the earliest time at which hoxa1 returned to the control levels (Fig. 7) intermingling the clutches of both studies. We observed high variability in the response to the RA manipulation among the individual clutches. Clutches with significant hoxa1 upregulation following RA addition showed limited downregulation in response to RA biosynthesis inhibition (DEAB; clutches C, L, J, K, H, I in Fig. 7). Clutches E, D, F, G, A, B (Fig. 7) showed the opposite response, i.e. robust hoxa1 change following RA knockdown, while the RA increase induced a mild response. The suppressors of RA signaling, e.g., cyp26a1 and dhrs3, showed significantly altered expression in clutches (C, L, J, K, H) with larger deviations in hoxa1 expression and a strong and extended response to the addition of RA. In contrast, the differential expression RA producers (e.g., aldh1a2 and rdh10) after RA manipulation was relatively mild and did not fully align with the deviation in hoxa1 expression (Fig. 7). Such a heterogeneous correlation between the extent of feedback regulation and clutch-to-clutch variation in hoxa1 expression was observed across the RA metabolic network (Supplementary Fig. S4).
Asymmetric robustness in response to increased versus decreased RA signaling
We sought to quantitatively rank the robustness of each clutch in an integrated manner based on the expression shift of multiple hox genes and the feedback response of the RA metabolic network genes. We pursued a trajectory-based approach in which the samples of all biological repeats (clutches) at all time points were projected onto the first three principal components based on the expression of hox genes (hoxa1, hoxa3, hoxb1, hoxb4, hoxd4) or the RA metabolic network genes. A principal curve was fit to the projected data, representing the trajectory in which the system evolves over time for all clutches combined. Each clutch was visualized separately along the principal trajectory, allowing us to compare the temporal evolution of deviations between treated samples and controls (Figs. 8A,B; Supplemental Fig. S5). As a multi-gene measure of the robustness of each clutch, the net absolute distance between the treated samples and controls was computed at all time points along the principal curve. Our results revealed a wide range of clutch-to-clutch variability in robustness as assessed by the integrated deviation from the control trajectory of multiple hox genes (Fig. 8C; Supplemental Fig. S5). The distance between treated samples and controls decreased over time in nearly all the clutches with significant differences in the time taken to close the gap (Fig. 8C). Clutches with low RA increase robustness showed larger distances between samples and controls along the hox gene trajectory at any time point (Fig. 8A, clutch C), compared to robust clutches with decreasing distances (Fig. 8A, clutch E). By contrast, DEAB treatment resulted in an opposite robustness pattern (Fig. 8B), with clutch C showing the least deviation and clutch E the widest deviation (Fig. 8B top row and Fig. 8C). Clutch J exhibited intermediate robustness responses to both the increase and the reduction of RA levels (Fig. 8A,B). The anti-correlated pattern of RA robustness suggests an asymmetric tradeoff in the gastrula embryo to counter changes in RA levels (Fig. 8E). The hox responses to changed RA levels revealed that all biological repeats aligned along a negative-slope diagonal and covered the whole range of responses (Fig. 8E). This result suggested the establishment of an RA robustness gradient among the clutches.
A similar analysis of the RA network component changes showed a scattered pattern among the biological repeats with no clear trend. The feedback regulatory response based on the RA metabolic network trajectory map (Figs. 8A,B bottom rows; Supplemental Fig. S5) was variable across clutches, with a characteristic pattern of reduced deviation over time in several clutches (Fig. 8D). There was no significant correlation in the net shift in feedback regulatory action between RA addition versus reduction (Fig. 8F). Interestingly, the scattering distribution depended on the direction of the RA manipulation. The range of clutch distribution for DEAB treatment suggests an immediate response to any RA reduction and an upper limit to this response (Fig. 8F). In contrast, the clutch distribution RA addition suggests that the response requires a threshold change to be activated and it exhibits also an upper threshold. These observations suggest a high sensitivity to RA reduction but a lower sensitivity to slight RA increase.
Clutch robustness variability affects the efficiency-efficacy of the response
We examined whether the asymmetric robustness to RA change direction correlates to the extent of the feedback regulatory effect (Fig. 9). We formulated an “efficiency-efficacy matrix” (Fig. 9A). The sections of this matrix delineate distinct possibilities based on all permutations of the robustness level based on the net shift in RA metabolic network, and the phenotypic outcome determined by the shift in the hox gene expression. High robustness (low net shift of hox genes) may be achieved by a mild (efficient) or a strong (effective) response in the RA metabolic network gene expression (lower quadrants in Fig. 8A). Similarly, the cases of low robustness (high net shift hox gene expression) may be accompanied by mild (inefficient) or strong (ineffective) gene expression changes in the RA metabolic network components (upper quadrants in Fig. 9A). In response to increased RA levels, half of the clutches were distributed within the efficient and effective zones (Figs. 9A,B; blue letters). Another five clutches exhibited a substantial shift in hox and RA network expression in increased RA, i.e. ineffective response. One clutch was located in an inefficient zone. By contrast, the distribution of clutches in response to RA reduction (DEAB) distributed between the efficient and inefficient quadrants, characterized by limited gene expression shift in the RA metabolic network while yielding a wide range of robustness in hox gene expression (Figs. 9A,B; green letters).
We examined if the clutches with similar levels of robustness and overall feedback regulation employ similar strategies for feedback regulation. Interestingly, we found that the identity of the differentially regulated genes in the RA metabolic network can vary between clutches co-localized in the efficiency-efficacy matrix (Figs. 9B,C). For example, in response to increased RA, clutches A, E and F showed similar differential regulation of dhrs3, aldh1a2, and cyp26a1, but diverged in stra6, sdr16c5, adhfe1, rdh13, aldh1a3, and cyp26c1 (Figs. 6E and 9C). Similarly, in response to reduced RA, clutches A, B and D showed downregulation of dhrs3 and cyp26a1, and upregulation of rbp1, suggesting feedback aimed at reducing enzymes active in RA reduction, and increasing import of retinol. However, there was high variability in the differential expression of aldh1a2 and aldh1a3 to increase RA production (Figs. 6E and 9B,D). Taken together, these results provide strong evidence that the early embryo is capable of robust maintenance of RA levels by mounting a range of network-wide feedback regulatory responses.
Discussion
Robustness of the retinoic acid signaling pathway
Signaling pathway robustness is a central characteristic of all regulatory networks to ensure signaling consistency and reliability overcoming changing environmental conditions or genetic polymorphisms. Gastrula stage Xenopus laevis embryos contain RA levels around the 100 nM - 150 nM range (Durston et al., 1989; Kraft and Juchau, 1992; Kraft et al., 1995). Importantly, to strong developmental defects, Xenopus embryos, and many other systems, are commonly treated with RA concentrations in the 1 μM to 10 μM range (Sive et al., 1990; Taira et al., 1994). To focus our study of robustness in RA signaling to the physiological range, we show that by increasing the RA levels by as little as 10 nM (about 10%) we could consistently induce expression changes in RA-regulated genes. Interestingly, embryo treatments with physiological RA concentrations exhibit very mild developmental malformations suggesting the induction of compensatory mechanisms to prevent abnormal gene expression (Hollemann et al., 1998; Reijntjes et al., 2005; Sive et al., 1990). We obtained similar results when we reduced the levels of RA by either blocking the biosynthesis (DEAB treatment) or by targeting this ligand for degradation (CYP26A1 overexpression). These observations are supported by multiple loss-of-function studies describing mild developmental malformations induced by RA signaling reduction (Blumberg et al., 1997; Hollemann et al., 1998; Janesick et al., 2014; Koide et al., 2001; Shabtai et al., 2018; Sharpe and Goldstone, 1997). Therefore, increased or decreased RA levels in the physiological range result in mild developmental defects, suggesting the activation of compensatory mechanisms. We show that one approach to overcome the robustness of RA signaling is to interfere with the metabolic/signaling network at multiple steps (DEAB+CYP26A1) to hamper its ability to efficiently elicit a feedback regulatory response.
We describe a systems biology approach to study the robustness of RA metabolism and signaling based on transient manipulation of ligand levels followed by temporal kinetic, transcriptome-wide analysis of the restoration of normal gene expression patterns. The large transcriptomic differences observed corresponded to normal expression changes resulting from the progression through embryogenesis. Alternatively, the treatments were inefficient, a possibility we could rule out by qPCR screening of all RNA samples for Hox expression changes prior to sequencing and subsequent computational analysis of the RNAseq data for individual gene expression shifts. The close clustering of the RA-manipulated (increased or decreased) and control samples in the PCA analysis at each time point further supports the robustness of RA signaling to maintain normal, non-teratogenic, target gene expression levels during early gastrulation.
The RA robustness response emerges from autoregulatory changes in the RA metabolic network
What is the mechanism activated to achieve RA robustness within the physiological range as observed in our study? The early qPCR analysis provided insights on the RA robustness mechanism showing that target genes exhibited abnormal expression levels at the end of the manipulation (t=0), and some target genes (hox) exhibited a speedy return to normal expression after 1.5 hours from the end of the treatment. Genes encoding RA metabolic network components also exhibited abnormal expression levels at t=0, but, their return to normal expression was delayed beyond the time required for the RA targets to reach normal expression. These observations suggest that the RA metabolic network is altered through an RA-dependent feedback regulatory mechanism to restore normal RA signaling and target gene expression. Multiple reports have described the RA-dependent regulation of individual RA metabolic components. Most studies show upregulation of enzymes suppressing or reducing RA signaling like CYP26A1, ADHFe1, and DHRS3, or downregulation of RA producers (anabolic enzymes) like RALDH2 and RDH10 as a result of RA treatment (Chen et al., 2001; Dobbs-McAuliffe et al., 2004; Fujii et al., 1997; Hollemann et al., 1998; Kam et al., 2013; Sandell et al., 2012; Shabtai et al., 2017; Sonneveld et al., 1998; Strate et al., 2009).
To understand the full extent of the kinetic transcriptomic response to transient RA manipulation, we analyzed the recovery kinetics by RNAseq. A discretized pattern analysis of kinetic patterns revealed that only 10 out of the 27 possible patterns were represented in any of the treatment or control samples, and only five patterns were exhibited by a substantial number of genes (n>100). These observations suggest that the RA-regulated genes, direct or indirect, can only exhibit a limited number of possible regulatory responses although intensity differences are possible.
To further understand the regulatory responses of RA targets, we performed a comparative analysis of the discretized patterns using our unbiased approach, COMPACT (Kuttippurathu et al., 2016). This analysis revealed that most genes (75.40%) responding to RA manipulation and recovery exhibited only a response to either increased or decreased RA levels. Only 48 genes out of 553 (8.67%) exhibited reciprocal responses to increased and decreased RA levels. Interestingly, many of these genes encode RA metabolism or signaling components to possibly maintain non-teratogenic RA levels. The temporal discrepancy in the return to normalcy between targets and network components and the network-wide response taking place at multiple levels provide an integrative view of the feedback regulation and robustness response. Initially, the response to RA changes will most probably be dealt by the actual enzymes and factors already present in the cell. In parallel, the same RA changes will elicit a transcriptional response which for increased RA should involve the upregulation of RA suppressor activities (DHRS3, ADHFe1, and CYP26A1), and the complementary downregulation of RA producing enzymes (RALDH2 and RDH10). This transcriptional response will exhibit a slight delay due to the transcription, translation, and post-translational modifications required.
We observed that some RA network components exhibit oscillatory behavior close to control expression levels, probably as a result of the fine-tuning of the RA signal. The fine-tuning of the RA signal levels coupled to the inherent delay in the transcriptional response could transiently result in the inversion of the overall signaling direction, and oscillatory transcriptional behavior. In a few instances such paradoxical observations of RA signaling outcome following RA manipulation, i.e. overcompensation, have been reported (D’Aniello and Waxman, 2015; D’Aniello et al., 2013; Lee et al., 2012; Rydeen et al., 2015). These observations identify a very dynamic feedback regulatory network continuously fine-tuning itself in response to perturbations.
Asymmetric response to increased and decreased RA levels
Our study provided new insights on the network responses to increased and decreased RA levels. One important conclusion is the observation that these responses are not symmetrically inverse but rather governed by different regulatory rules. Following a trajectory analysis, we conclude an RA auto-regulatory asymmetry based on:
First, the activation of a response to reduced RA signaling after very slight reduction, while the response to increased RA is only activated above a threshold. Also, the response to reduced RA reaches a lower upper threshold than the response to increased RA. Notably, the network responses show different thresholds depending on the direction of the RA manipulation suggesting a high sensitivity to RA reduction but a lower sensitivity to slight RA increase.
Second, the alignment of the robustness responses, hox changes, fitted a diagonal with a negative slope. This negative slope suggested that while a clutch might very efficiently deal with increased RA, i.e. high robustness, the same clutch struggles to compensate for a reduction in RA, i.e. low robustness. The inverse situation, and also more “balanced” clutches were observed suggesting an asymmetric response to RA manipulations.
Third, different clutches with similar robustness levels to increased or reduced RA levels can mount robustness responses by incorporating different components of the RA network. The transcriptomic and HT-qPCR analyses allowed us to perform a detailed determination of the network components comprising a robustness response to uncover its mechanism. The RA metabolic network includes multiple enzymes whose biochemical activity overlaps, e.g., ALDH1A1 (RALDH1), ALDH1A2 (RALDH2), AND ALDH1A3 (RALDH3) enzymes that oxidize retinaldehyde to produce RA (Ghyselinck and Duester, 2019; Kedishvili, 2013; Kedishvili, 2016; Shabtai et al., 2016). The enzymatic efficiencies and expression patterns including location, timing, and intensity may be different across the enzymes (Blentic et al., 2003; Chen et al., 2001; Lupo et al., 2005; Romand et al., 2004; Shabtai et al., 2016). The enzymes CYP26A1, B1, and C1, or DHRS3, ADHFe1, and RDH13 function to prevent excessive RA signaling (Belyaeva et al., 2008; Belyaeva et al., 2017; Hollemann et al., 1998; Shabtai et al., 2017; Sonneveld et al., 1998). These redundant enzymes and factors can help establish a robustness response using different components.
Finally, genetic polymorphisms probably explain in part the different responses between clutches of the outbred Xenopus laevis (Savova et al., 2017). Formally, technical issues could account for some variability. The six additional clutches analyzed by HT-qPCR showed the same variability supporting the involvement of genetic polymorphisms. We mined the Savova et al. (2017) data identifying multiple polymorphisms in RA network components (Supplemental Table S4), supporting the potential contribution of genetic variability in the differential robustness.
Retinoic acid signaling changes due to environmental changes and disease risk
While the RA exposure induces dramatic developmental malformations, treatment of the same embryos with ROL or RAL, the RA precursors, requires higher concentrations (70-100X and 5X respectively) to induce similar defects (Durston et al., 1989; Yelin et al., 2005). Members of the CYP26 family should partially neutralize the RA added or produced (Dobbs-McAuliffe et al., 2004; Hollemann et al., 1998; Sakai et al., 2001). The main difference is that ROL or RAL have to go through RA biosynthesis while RA is already the final active ligand. The oxidation of rol to RAL is a reversible reaction that can reduce substrate availability for the ALDH enzymes, while treatment with RA can only be countered through inactivation by CYP26 enzymes. Therefore, the reduced teratogenicity of the precursors could be the result of reduced conversion to RA as part of the feedback regulation of this network, i.e. robustness, while RA treatment is more restricted in its robustness response.
We obtained evidence of the RA signaling robustness employing several experimental approaches. We observed an efficient response to RA manipulation such that the transcriptomes of treated embryos were not significantly different from controls. All our experiments were performed by either partially inhibiting the endogenous levels of RA, or by slightly increasing (~10% increase) the physiological RA content. Under these conditions, the RA robustness of the embryo efficiently normalizes the transcriptome via regulatory feedback. Based on the comparative analysis of the 12 clutches (genetic backgrounds), we can suggest that robustness efficiency will have a threshold beyond which it will become ineffective in restoring normal RA signaling. Our clutch analysis suggests that this threshold might be strongly dependent on genetic polymorphisms affecting enzymatic activity or gene expression parameters. In support, a threshold or toxicological tipping point for RA signaling was recently described in a cell-based model (Saili et al., 2019). The clutch analysis also showed that the RA network response is also dependent on genetic variability like promoter polymorphisms. Then, the developmental malformations arising from environmental insults on RA signaling largely depend on genetic polymorphisms which will determine the efficiency and threshold of the response and the actual network components comprising such a response.
Materials and Methods
Embryo culture
Xenopus laevis frogs were purchased from Xenopus I or NASCO (Dexter, MI or Fort Atkinson, WI). Experiments were performed after approval and under the supervision of the Institutional Animal Care and Use Committee (IACUC) of the Hebrew University (Ethics approval no. MD-17-15281-3). Embryos were obtained by in vitro fertilization, incubated in 0.1% MBSH and staged according to Nieuwkoop and Faber (1967).
Embryo Treatments
all-trans Retinoic acid (RA), Dimethyl sulfoxide (DMSO), and 4-Diethylaminobenzaldehyde (DEAB), were purchased from Sigma-Aldrich (St. Louis, Missouri). Stock solutions of RA, and DEAB, were prepared in DMSO. Two-hour treatments of 10 nM RA, or 50 μM DEAB, were initiated during late blastula (st. 9.5) and terminated at early gastrula (st. 10.25) by three changes of 0.1% MBSH and further incubation in fresh 0.1% MBSH for the desired time.
Expression analysis
For each sample, 5-10 staged embryos were collected and stored at −80°C. RNA purification was performed using the Bio-Rad Aurum Total RNA Mini Kit (according to the manufacturer’s instructions). RNA samples were used for cDNA synthesis using the Bio-Rad iScriptTM Reverse Transcription Supermix for RT-qPCR kit (according to the manufacturer’s instructions). Quantitative real-time RT-PCR (qPCR) was performed using the Bio-Rad CFX384 thermal cycler and the iTaq Universal SYBR Green Supermix (Bio-Rad). All samples were processed in triplicate and analyzed as described previously (Livak and Schmittgen, 2001). All experiments were repeated with six different embryo batches. qPCR primers used are listed in (Supplemental Table S2).
RNASeq data analysis
Sequencing was performed at the Thomas Jefferson University Genomics Core using Illumina HiSeq 4000. Reads were mapped to the genome using the Xenopus laevis 9.1 genome with STAR alignment and a modified BLAST and FASTQC in the NGS pipeline (STAR average mapped length: 142.34). Annotation of the mapped sequences using verse identified 31535 genes. Raw counts were further filtered for non-zero variance across all samples resulting in 31440 scaffold IDs.
High throughput qPCR
cDNA samples were directly processed for reverse transcriptase reaction using SuperScript VILO Master Mix (Thermo Fisher Scientific, Waltham, MA), followed by real-time PCR for targeted amplification and detection using the Evagreen intercalated dye-based approach to detect the PCR-amplified product. Intron-spanning PCR primers were designed for every assay using Primer3 and BLAST for 24 genes from Retinoic Acid metabolism and target pathway (Supplemental Table S2). The standard BioMark protocol was used to preamplify cDNA samples for 22 cycles using TaqMan PreAmp Master Mix as per the manufacturer’s protocol (Applied Biosystems, Foster City, CA, USA). qPCR reactions were performed using 96.96 BioMark Dynamic Arrays (Fluidigm, South San Francisco, CA, USA) enabling quantitative measurement of multiple mRNAs and samples under identical reaction conditions. Each run consisted of 30 amplification cycles (15 s at 95°C, 5 s at 70°C, 60 s at 60°C). Ct values were calculated by the Real-Time PCR Analysis Software (Fluidigm). Samples were run in triplicate for the 24 genes. The primers in the first pre-amplification group selectively bind for the L homoeologues of the genes. In the second pre-amplification group, primers were selected for the S homoeologues only. Third pre-amplification group binds to all 24 genes. In order to remove the technical variability caused due to L and S gene homoeologues, Ct value for each gene in a sample was selected as the median value of the three pre-amplification runs. Relative gene expression was determined by the ΔΔCt method. Gapdh was used as a housekeeping gene for normalization of the data.
Data normalization and annotation
Data analysis on raw genes count was performed using the R statistical analysis tool version 3.6.0 on a 64-bit Windows platform. For the RNA-seq data, the raw genes counts were first converted into log2-transformed values using the “regularized log (rlog)” transformation from the DESeq2 package, which minimizes differences between samples for rows with small counts (Love et al., 2014). The gene expression data was then normalized across samples against the experimental variation and batch effects using COMBAT method in R using a non-parametric adjustment (Johnson et al., 2007). Following batch correction, the gene list was filtered for a minimum expression threshold to remove genes with normalized counts less than 5 across all 72 samples. The expression data for the remaining genes was normalized using quantile normalization. RNA-Seq transcript/array probes IDs were transformed to Official Gene Symbol using merged list from 3 sources: the Xenopus laevis scaffold-gene mapping, DAVID Bioinformatics Resource 6.8 (Huang et al., 2009) or AnnotationData package “org.Xl.eg.db” maintained by Bioconductor (Carlson, 2017). The original scaffold IDs were retained along with the Official Gene Symbols for cross-reference purposes.
Differential gene expression analysis
The normalized data was analyzed using an Empirical Bayes Statistical (eBayes) model that considered the following two variables and their interactions: (1) Treatment (Control, RA, DEAB) and (2) Time post treatment-washout (t = 0, 1.5, 3, 4.5h). Differentially expressed genes were identified based on statistically significant effects of Time, Treatment or an interaction between these two factors. P-values were adjusted for multiple testing using topTable from limma (Ritchie et al., 2015) package in R (q ≤ 0.05). The significance-filtered differential gene expression data was used in an established Principal Component Analysis (PCA) approach using the prcomp function implemented in R. The samples were annotated based on a combination of treatment and time, yielding 12 distinct sample groups. For each of the selected PC, expression of 100 top positively-loaded and 100 top negatively-loaded genes was visualized using a heat map.
Dynamic pattern analysis and COMPACT analysis
First, for each time point for each of the treatment conditions, the gene expression data for all six clutches (A,B,C,D,E,F) was averaged. Within treatment groups, RA, DEAB, Control, the gene expression data at time points t=1.5, 3, 4.5h was normalized by subtracting the corresponding ‘t=0h’ group. This average differential gene expression data was then discretized to three levels (+1, 0, −1) based on a fold-change threshold [±2 (up, no or down-regulation)]. Within the three treatment groups, this discretization yielded a dynamic response pattern vector for each gene, encoded by one of 27 (3 levels^3 time-points) possible ordered sets. Counts of genes in each treatment group that follow each of the 27 * 27 (=729) possibilities were compared. Functional enrichment analysis was performed for geneset in various dynamic pattern vectors in the Control conditions, using functions enrichGO and simplify from the R package “clusterProfiler” (Yu et al., 2012).
COMPACT analysis of RA and DEAB after normalizing to Control
First, for each time point for each of the treatment conditions, the gene expression data for all six clutches (A,B,C,D,E,F) was averaged. Within treatment groups, RA and DEAB, the gene expression data were normalized to Control at each time point by subtracting the expression of the Control group at the corresponding time point. This Control-normalized differential gene expression data for both treatment groups was then discretized to three levels (+1, 0, −1) based on a fold-change threshold [±log2(1.3) (up, no or downregulation)]. This discretization yielded a dynamic response pattern vector for each gene, encoded by one of 81 (3 levels^4 time-points) possible ordered sets. Subsequently, RA and DEAB groups were compared to count the number of genes corresponding to each of the 81 * 81 (=6561) possibilities; to create an 81 × 81 matrix representing the comparative dynamic response pattern counts (COMPACT) (Kuttippurathu et al., 2016). For a given COMPACT matrix of comparative conditions RA(vs. Control) and DEAB (vs. Control), the element at the ith row and jth column of the matrix contains the number of genes that show an ‘i’th pattern in DEAB and ‘j’th pattern in RA. For a coarse-grained version of the detailed 81×81 COMPACT, pattern-vector counts for each treatment group were further aggregated based on the first time-point, yielding 9 groups of pattern vectors per treatment group. The pair of treatment groups (RA and DEAB) was then compared to count the number of genes corresponding to each of the 9 * 9 (=81) possibilities.
RA network map and visualization
A schematic representation for the position and functioning of the genes involved in the RA biosynthesis, metabolism, translocation, and transcription was formulated from the literature. For each time point for each treatment, expression value for each gene was mapped to the corresponding label in the schematics using a color scale.
Gene correlation and clustering analysis -Gene expression data was analyzed for both with and without Control-normalization using Weighted Gene Coexpression Network Analysis (WGCNA) (Langfelder and Horvath, 2008) to identify modules of genes with highly correlated differential expression. We used a soft threshold value of 9 (8 for Control normalized data) to identify the initial gene coexpression modules, followed by a dissimilarity threshold of 0.25 to merge the initial modules into the final set of gene coexpression modules. Identified modules were further correlated with the traits (batch, time, treatment).
Robustness score calculation and principal curve-based trajectory analysis
RNA-seq expression data of (clutches: A,B,C,D,E,F) and additional normalized qPCR expression data (clutches: G,H,I,J,K,L) were independently genewise Z-score transformed. Transformed data was then combined to result into 144 samples (12 clutches * 3 treatments * 4 time points). We further selected two subsets of genes: 1) “RA metabolism genes” [aldh1a2.L, aldh1a3.L, crabp2.L, crabp2.S, cyp26a1.L, cyp26a1.S, cyp26c1.L, cyp26c1.S, dhrs3.L, rbp1.L, rdh10.L, rdh10.S, rdh13.L, rdh14.L, sdr16c5.L, stra6.L]. This set represents the feedback regulatory mechanism utilized in response to the RA levels perturbations. 2) “HOX genes” [hoxa1.L, hoxa1.S, hoxa3.S, hoxb1.S, hoxb4.S, hoxd4.L] representing the phenotypic outcome from the treatment. For each gene set, the PCA scores for the combined data was first calculated using R function prcomp and then the scores from the first three principal components (PC1, PC2 and PC3) are used to learn a 3-dimensional principal curve using the function principal_curve from R package princurve. Two sets of points are specified for the ‘start’ parameter for this function, which determines the origin and direction of the curve: (1) The centroid of the 0h-Control samples from all twelve clutches, and (2) centroid of all the remaining 132 samples. For each geneset (RA metabolism genes and HOX genes), for each clutch (A-L), for each treatment (RA or DEAB), a “net absolute expression shift” can be calculated as the sum of distances along the principal curve, of the treatment samples from the corresponding control samples: where, is the arc-distance from the beginning of the curve, for the “RA treatment samples” at time point “t”, for clutch “cl” for the principal curve learned for the geneset “Hox genes” (Returned as the parameter “lambda” from the principal curve function). The expression shift calculation allows ranking and sorting clutches based on ‘HOX shift’ and ‘RA Network shift’ . A higher HOX shift for a clutch for a treatment indicates that the clutch is more robust to that particular treatment/perturbation. Furthermore, clutches were mapped onto the conceptual map across the spectrum of “HOX shift” and “RA Network shift” divided into hypothetical quadrants of Effective or Ineffective regulation, or Efficient or Inefficient regulation.
Availability of supporting data/additional files
The raw and normalized datasets for the RNAseq and HT-qPCR data are available online as Gene Expression Omnibus datasets via SuperSeries GSE154408 containing RNAseq data: GSE154399 and HT-qPCR data: GSE154407.
Funding
RV acknowledges financial support for the project from the National Institute of Biomedical Imaging and Bioengineering grant U01 EB023224 and from the Department of Pathology, Anatomy, and Cell Biology, Thomas Jefferson University. AF acknowledges financial support from the Israel Science Foundation (grant 668/17) and the Wolfson Family Chair in Genetics. RV and AF acknowledge the Pilot Funding grant from Thomas Jefferson University and The Hebrew University of Jerusalem collaborative research program.
Author Contributions
R.V. and A.F. conceived and supervised the study and designed the experiments and analysis methodology. L.B-K., M.G., K.K., and A.F. performed embryo experiments and real-time PCR assessment and developed the figures. A.B. performed the initial analysis of the RNAseq data. S.A. conducted the high-throughput PCR validation of RNAseq results. M.P. conducted the analysis of transcriptomics and HT-qPCR data and performed the network and trajectory analyses and developed the figures. M.P., R.V., and A.F. interpreted the results and drafted the manuscript.
Ethics Declarations
The authors declare no competing interests.
Figure Legends
Footnotes
Review Codes for online datasets GEO SuperSeries GSE154408: cryfqygubnszryt https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE154408
Revised abstract, introduction, results, and discussion. Updated figures and supplemental files.
https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE154408
Abbreviations Used
- COMPACT
- Comparative Matrix of Pattern Counts
- DEAB
- 4-diethylaminobenzaldehyde
- PCA
- Principal Component Analysis
- PCR
- Polymerase Chain Reaction
- RA
- retinoic acid
- WGCNA
- Weighted Gene Coexpression Network Analysis